Knowledge Extraction from Local Domain Corpora Using Neurosymbolic Representation
Dr. Houbing Song, Department of Information Systems.
Large language models score well on knowledge-intensive QA benchmarks but at high computational cost and without auditable reasoning. We present a methodology for constructing neurosymbolic parallel corpora from unstructured text: line-aligned symbolic representations stored as sparse tensors T[s, f] and queried via SQL. On the OOLONG benchmark (6,072 questions over Critical Role tabletop gaming transcripts), programmatic symbolic retrieval via the sparse tensor scores 0.9939, and a Mistral-7B adapter trained on these reasoning traces scores 0.8453. Both outperform Gemini-2.5-Pro (0.5295), GPT-5 (0.4700), and Claude-Sonnet-4 (0.3675) by 60–88% relative. A semantic-only baseline using identical answerer code scores 0.0322. We also show that reasoning patterns developed over the symbolic substrate can serve as training data for a smaller domain-specific model that reaches comparable performance through grounded inference rather than parametric memorization. The methodology transfers to new domains by changing only the vocabulary library.
Dr. Houbing Song, Department of Information Systems.
The interplay between AI and cybersecurity introduces new opportunities and challenges in the cybersecurity of AI as well as AI for cybersecurity. However, operations and configurations of AI cyberinfrastructure (CI) with a security mindset are rarely covered in the typical AI curriculum. To fill this gap, this project intends to develop hands-on training materials and provide mentored training for the current and future research workforce in engineering and science-related disciplines. By transforming and integrating training materials into a course curriculum, this project aims to train potential cyberinfrastructure professionals in the CI community at large to handle AI with and for cybersecurity. This project has the potential to develop the research workforce in operating AI cyberinfrastructure with a security mindset to meet the national and economical needs and priorities of CI advancement.